// Yapay Zeka

OVHcloud announced the open-source model Olala

OVHcloud introduced its first large language model family, Olala, along with Olala mini 7A1B, which runs with just 8 GB of GPU memory. The model will be released as open source on Hugging Face under the Apache 2.0 license.

French cloud provider OVHcloud introduced its first large language model family, Olala, and the first member of that family, Olala mini 7A1B, on November 17, 2026. According to the company, the model can run on a GPU with just 8 GB of memory and will be released as open source on Hugging Face under the Apache 2.0 license. The official unveiling will take place at the OVHcloud AI Summit on November 19, 2026. The release package includes model weights, a pre-trained base model, and compliance documents for the European Union AI Act.

Hybrid attention and low hardware requirements

Olala mini has a 7 billion-parameter architecture, but only 1 billion parameters are activated for each request. The active parameters are distributed across 256 experts per layer, so each request uses only a small part of the network instead of the entire network. The model’s hybrid attention design combines Differential-TPA with Mamba-3 and uses the BF16 format. OVHcloud reports speeds of 160 tokens per second on an RTX A4500 and 50 tokens per second on a standard CPU without a GPU. The model’s current context window is 32,000 tokens.

Trained on European supercomputers

The model was trained from scratch on three European public supercomputers: Jupiter, MeluXina, and Jean Zay. The training infrastructure used GH200, A100, and H100 GPUs. OVHcloud says the public support amounted to 7 million euros in compute time and was decisive for the project. The effort also brought together the company’s AI team and led to the creation of the Agentic Lab called OVHai LLM. That positions Olala mini not as a one-off training experiment, but as the beginning of an ongoing model family.

Qwant is the first customer, and the roadmap is expanding

Search engine Qwant has started using Olala mini in its flash-answer feature. OVHcloud says responses are generated 20% faster and operating costs have dropped to one-third of the previous model. According to the roadmap, version 1.1, due at the end of 2026, will expand the context window to 128,000 tokens and improve mathematical reasoning. Version 2.0, which is planned to exceed 100 billion parameters, will specialize in agentic AI for business, cloud DevOps, and cybersecurity. The company also aims to gradually bring Olala into its own products and operational tools.

OVHcloud CEO Octave Klaba said the company needs to develop models both for its own needs and for customer demand: “In the coming years, we cannot blindly rely on the generosity of some labs to open-source their LLMs.” Klaba added that training expertise is becoming more widespread and costs have fallen 20-fold in three years, which he sees as the path to building profitable models.

Source: Unite.AI

// JOIN THE CONVERSATION

Share your experience

Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.

// KEEP READING

One more note.

All posts ↗︎